Upload
vonguyet
View
221
Download
0
Embed Size (px)
Citation preview
CNR-IEIIT
Information-Based Complexity for Systems and Control: The Probabilistic Setting
Consiglio Nazionale delle Ricerche
Hara Fest at The University of Tokyo @RT 2012
Roberto Tempo
CNR-IEIITConsiglio Nazionale delle Ricerche
Politecnico di [email protected]
CNR-IEIIT Happy Birthday to Shinji!
Hara Fest at The University of Tokyo @RT 2012
CNR-IEIIT Acknowledgments
Research on the probabilistic setting of Information-Based Complexity is joint work with
Fabrizio Dabbene
Hara Fest at The University of Tokyo @RT 2012
Mario Sznaier
CNR-IEIIT
This talk deals with problems which are solvableonly approximately because information is partial
or contaminated
This Lecture - 1
Hara Fest at The University of Tokyo @RT 2012
o co ed
J.F. Traub, G.W. Wasilkowski, H. Wozniakowski, “Information-Based Complexity (IBC)”, 1988
CNR-IEIIT
This talk deals with problems which are solvableonly approximately because information is partial
or contaminated
This Lecture - 2
Hara Fest at The University of Tokyo @RT 2012
o co ed
Objectives:o derive optimal algorithmso compute approximation error o analyze computational complexity
CNR-IEIIT
This talk deals with problems which are solvableonly approximately because information is partial
or contaminated
This Lecture - 3
Hara Fest at The University of Tokyo @RT 2012
o co ed
Different settings:o worst-case (classical)o probabilistic (new)
CNR-IEIIT This Lecture - 4
Different areas may be covered, e.g. information theory,complexity, signal processing, numerical analysis, …
IBC applications: integration problems, solutions ofnonlinear equations, etc
Hara Fest at The University of Tokyo @RT 2012
We are interested in systems and control and in thederivation of optimal algorithms for specific applications
CNR-IEIIT Spaces and Operators
problem element (unknown)Xx
Hara Fest at The University of Tokyo @RT 2012
CNR-IEIIT Spaces and Operators
Xx
exampleconsider input-output pair
problem element (unknown)
Hara Fest at The University of Tokyo @RT 2012
consider input-output pair ξ(x, t) of a dynamic system
with given basis φi(t)
T1
ξ( , ) φ ( ) ( )ni ii
x t x t t x
CNR-IEIIT Spaces and Operators
Xx
problem element (unknown)
Hara Fest at The University of Tokyo @RT 2012
Y
data
I
I x
CNR-IEIIT Spaces and Operators
Xx
problem element (unknown)
Hara Fest at The University of Tokyo @RT 2012
Y
I
I xy = I x + q
data
CNR-IEIIT Spaces and Operators
Xx
information operatorand measurements
problem element (unknown)
Hara Fest at The University of Tokyo @RT 2012
Y
I
I xy = I x + q
and measurementsm n noisy measurements of ξ(x, t) are available for
t1 < t2 < · · · < tmT
1[ ( ) ( )]my t t x q
data
CNR-IEIIT Spaces and Operators
Xx ZS
solution spaceproblem element (unknown)
z = S x
Hara Fest at The University of Tokyo @RT 2012
Y
I
I xy = I x + q
data
CNR-IEIIT Spaces and Operators
Xx ZS
solution spaceproblem element (unknown)
z = S x
Hara Fest at The University of Tokyo @RT 2012
Y
I
I xy = I x + q
solution operatorestimate future values of ξ(x, t) for tm+1 < · · · < tm+s
T1[ ( ) ( )]m m sx t t x S
data
CNR-IEIIT Spaces and Operators
Xx ZS
solution space
zz = S x
problem element (unknown)
Hara Fest at The University of Tokyo @RT 2012
Y
I
I xy = I x + q
A algorithmprovides an estimate
of z = Sxzdata
CNR-IEIIT Ingredients
Problem element x X = Rn with prior information K
Information operator (linear) I: X → Y = Rm (m n)
Information I x corrupted by noise q
Hara Fest at The University of Tokyo @RT 2012
Data y = I x + q
Bounding set Q for q
Solution operator (linear) S: X → Z = Rs (n s)
Algorithm (nonlinear) A: Y → Z
CNR-IEIIT Problem Element x
Problem element (unknown) x K X
K represents prior information (if available)
Hara Fest at The University of Tokyo @RT 2012
CNR-IEIIT Problem Element x
Problem element (unknown) x K X
K represents prior information (if available)
example
Hara Fest at The University of Tokyo @RT 2012
consider input-output pair (x, t) of a dynamic system
T1
ξ( , ) φ ( ) ( )ni ii
x t x t t x
: 0, 1, 2, ,iK x X x i n
CNR-IEIIT Noise q and Bounding Set Q
We assume that
q Q = q: ||q|| Rm
where ||.|| is the lp norm
Hara Fest at The University of Tokyo @RT 2012
CNR-IEIIT Approximation Error
We study the approximation error
where ||.|| is the lp norm
|| ( ) ||x yS A
Hara Fest at The University of Tokyo @RT 2012
w e e ||.|| s e p o
CNR-IEIIT Consistency Set
Consistency set is defined as
1( ) : there exists : = +y x K q Q y x q I I
Hara Fest at The University of Tokyo @RT 2012
CNR-IEIIT Consistency Set I-1(y)
X I-1(y)
ZSS I-1(y)
z
Hara Fest at The University of Tokyo @RT 2012
Y
I
yQ A
CNR-IEIIT
Hara Fest at The University of Tokyo @RT 2012
Worst-Case Setting
CNR-IEIIT Objective for Worst-Case Setting
Objective: construct an algorithm A (nonlinear)= A(y) of z = S x
and compute the worst-case radius rwc(A, y)z
Hara Fest at The University of Tokyo @RT 2012
o prior information x K X (if available)
o data y = I x + q Yo bounding set Q = q: ||q|| Rm
CNR-IEIIT
Worst-Case Radius and Optimal Algorithm
Given data y and the algorithm A, the worst-case radiusis defined as
1
wc
( )( , ) max || ( ) ||
x yr y x y
IA S A
Hara Fest at The University of Tokyo @RT 2012
Given y a worst-case optimal algorithm is defined as
is the worst-case radius of the optimal algorithm
( )x yI
1
wc wc wco o
( )( ) ( , ) inf max || ( ) ||
x yr y r y x y
A IA S A
wcoA
wco ( )r y
CNR-IEIIT Error of the Algorithm
X I-1(y)
ZS
z
z
S I-1(y)
Hara Fest at The University of Tokyo @RT 2012
Y
I
yQ A
z
CNR-IEIIT Worst-Case Optimal Algorithm
X I-1(y)
ZS zS I-1(y)
z
Hara Fest at The University of Tokyo @RT 2012
Y
I
yQ A
CNR-IEIIT
Optimal Algorithms and Chebychev Center
X I-1(y)
ZS ˆcz
|||| overbounding of S I-1(y)
Hara Fest at The University of Tokyo @RT 2012
Y
I
yQis the Chebychev
center of the set S I-1(y)is worst-case optimal
ˆcz
ˆcz
CNR-IEIIT
Hara Fest at The University of Tokyo @RT 2012
Probabilistic Setting
CNR-IEIIT
In the probabilistic setting we reduce theconservatism of the worst-case setting
at the expense of a “small” risk
Probabilistic Setting: Conservatism Reduction
Hara Fest at The University of Tokyo @RT 2012
violation functionvo(r) shows how theprobabilistic riskε (0,1) changesas a function ofthe radius r
worst-caseradiusprobabilistic
radiusrisk
CNR-IEIIT Random Noise
Assume that noise q is a random vector with uniform pdfU [Q] and support set Q
Hara Fest at The University of Tokyo @RT 2012
CNR-IEIIT Uniform Density U [Q]
Univariate uniform density
b
1/(b-a)[ , ]a bU
Hara Fest at The University of Tokyo @RT 2012
Multivariate uniform density U [Q]
1 if
vol( )0 otherwise
q QQQ
U
a b
CNR-IEIIT Objective for Probabilistic Setting
Objective: construct an algorithm A (nonlinear)= A(y) of z = S x
and compute the probabilistic radius rpr(A, y, ε)z
Hara Fest at The University of Tokyo @RT 2012
o probabilistic risk ε (0,1)o prior information x K X (if available)
o data y = I x + q Yo random vector q with uniform pdfo bounding set Q = q: ||q|| Rm
CNR-IEIIT Probabilistic Radius
Given data y, accuracy ε (0,1) and the algorithm A, theprobabilistic radius is defined as
1
pr
: ( ) ε ( )\( , ,ε) inf max || ( ) ||
x yr y x y
IA S A
Hara Fest at The University of Tokyo @RT 2012
Remark: we replace the consistency set with
We discard sets Ω of measure μ(Ω) < ε Probabilistic radius is smaller than worst-case radius
( )y
1( )yI
1( ) \y I
CNR-IEIIT Consistency Set I-1(y)
X I-1(y)
ZS
S I-1(y)
z
Hara Fest at The University of Tokyo @RT 2012
Y
I
yQ A
CNR-IEIIT Set
X I-1(y)\Ω
ZS z
1( ) \y I
S I-1(y)
Hara Fest at The University of Tokyo @RT 2012
Y
I
yQ A we discard sets Ωof measure μ(Ω) < ε from consistency set
CNR-IEIIT Probabilistic Optimal Algorithm
Given data y and risk ε (0,1) a probabilistic optimalalgorithm is defined as
1
pr pr pro o : ( ) ε ( )\
( ,ε) ( , ,ε) inf inf max || ( ) ||x y
r y r y x y
A I
A S A
proA
Hara Fest at The University of Tokyo @RT 2012
is the probabilistic radius of optimal algorithm
( ) ( )\x y I
pro ( ,ε)r y
CNR-IEIIT Measure of Sets
Theorem: Let q ∼ U [Q] and K = Rn. Then, for any y ∈ Y
o is uniform 1( )x y I
1 1( ) and ( )y y I S I
Hara Fest at The University of Tokyo @RT 2012
o is log-concave
This result also holds when Q is a convex set
def log-concave:
1( )z x y S S I
1(1 )A B A B
CNR-IEIIT
Violation Function and its Computation
Hara Fest at The University of Tokyo @RT 2012
Violation Function and its Computation
CNR-IEIIT Optimal Violation Function
Given r > 0 and A, define a violation function
1( , ) ( ) : || ( ) ||v r x y x y r A I S A
Hara Fest at The University of Tokyo @RT 2012
Given r > 0, the optimal violation function vo(r) is
1o ( ) inf ( , ) inf ( ) : || ( ) ||v r v r x y x y r
A AA I S A
CNR-IEIIT Chance-Constrained Optimization
Computation of the probabilistic radius of the optimal algorithm
b f l d h i d bl
pr pr pro o( ,ε) ( , ,ε)r y r y A
Hara Fest at The University of Tokyo @RT 2012
may be reformulated as a chance-constrained problem
For any ε (0,1), we can compute solvinga one-dimensional optimization problem in r > 0
pr pro o( , ,ε) min : ( ) εr y r v r A
pr pro( , ,ε)r yA
CNR-IEIIT
Properties of Optimal Violation Function
Theorem: Let q ∼ U [Q], K = Rn and S = identity. Forfixed r > 0, optimal violation is given by
1
o 1
vol ( ) \ :|| ||( ) inf
l ( )cy x x x r
v r
I
I
Hara Fest at The University of Tokyo @RT 2012
This optimization problem is quasi-convex for all well-defined xc (i.e. non-zero volume)
Optimal violation vo(r) is right-continuous and non-increasing for r > 0
o 1( )
vol ( )cx yI
CNR-IEIIT Optimal Violation
I-1(y)X :|| ||x x x r
proof based on the Brunn-Minkovski inequalityfor intersection of convex sets
Hara Fest at The University of Tokyo @RT 2012
X xc :|| ||cx x x r
1
o 1
vol ( ) \ :|| ||( ) inf
vol ( )c
c
x
y x x x rv r
y
I
I
CNR-IEIIT Optimal Violation
I-1(y)
X
Hara Fest at The University of Tokyo @RT 2012
X xc :|| ||cx x x r
1
o 1
vol ( ) \ :|| ||( ) inf
vol ( )c
c
x
y x x x rv r
y
I
I
CNR-IEIIT Extensions and Algorithms
Extensions for non-identity solution operator S
Computation of optimal violation vo(r) requirescomputation of
V l i f l i NP h d 1vol ( ) \ :|| ||cy x x x r I
Hara Fest at The University of Tokyo @RT 2012
Volume computation of polytopes is NP-hard
Two relaxation algorithms:
o Hard - deterministic (SDP-based)
o Soft - probabilistic (randomized-based)
CNR-IEIIT Conclusions
This probabilistic setting is completely new withinsystems and control
Applications:
l i l id ifi i i i
Hara Fest at The University of Tokyo @RT 2012
o classical: identification, estimation, …
o modern: PageRank computation in Google
CNR-IEIIT
(modern): PageRank computation in Google
H. Ishii and R. Tempo“Distributed Randomized
example
Conclusions
* *(1 )( ) mx m H U x
Hara Fest at The University of Tokyo @RT 2012
Algorithms for thePageRank Computation,”IEEE TAC, 2010
x* is PageRank, m=0.15, nis the number of pages, H is
the hyperlink matrix, ∆ represents link failures and
U is a rank-one matrix
(1 )( )x m H U xn
CNR-IEIIT Main References
F. Dabbene and R. Tempo, “Probabilistic and Randomized Toolsfor Control Design,” The Control Handbook (W. S. Levine Ed.),Taylor & Francis, 2010
G. Calafiore, F. Dabbene and R. Tempo “Research onProbabilistic Design Methods,” Automatica, 2011
Hara Fest at The University of Tokyo @RT 2012
g , , R. Tempo, G. Calafiore and F. Dabbene, “Randomized
Algorithms for Analysis and Control of Uncertain Systems,”Springer-Verlag, London, 2005 (second edition in preparation)
http://staff.polito.it/roberto.tempo/